smollm3¶
SmolLM3 on vLLM, through its transformers backend (SmolLM3ForCausalLM -> TransformersForCausalLM).
vLLM has no SmolLM3 of its own: it runs transformers' modules, so the tree
and the block are transformers' (see nnterp.components.vllm.transformers_backend),
with Llama's names, the stream [1, tokens, hidden_size] and the engine's
attention layer mounted as the attention's attn child. The block is
Llama's, pre-norm with the residual added in the block, so the
contributions are the modules' outputs. Every fourth block's attention
skips the rotary embedding (no_rope_layers); the queries and keys are
served as the attention layer receives them, rotated or not.